Mechanism equivalence between closed-source frontier models and open-weight models

Determine whether the activation-space directions identified in open-weight models Llama 3.1 and Qwen 2.5-32B are the same mechanisms that produce the behavioural signatures of extended self-examination observed in closed-source frontier models Claude Opus 4.5 and ChatGPT 5.2.

Background

Behavioural signatures of extended self-examination were established on closed-source frontier models (Claude, ChatGPT, Grok), while mechanistic direction identification and steering were performed on open-weight models (Llama, Qwen).

The authors explicitly note that they cannot directly verify whether the directions found in open-weight models are the same mechanisms producing behavioural signatures in the closed-source models, due to lack of activation access.